Combined Single Cell Transcriptome and Surface Epitope Profiling Identifies Potential Biomarkers of Psoriatic Arthritis and Facilitates Diagnosis via Machine Learning

• Published in: Frontiers in immunology Vol. 13; p. 835760
• Main Authors: Liu, Jared, Kumar, Sugandh, Hong, Julie, Huang, Zhi-Ming, Paez, Diana, Castillo, Maria, Calvo, Maria, Chang, Hsin-Wen, Cummins, Daniel D., Chung, Mimi, Yeroushalmi, Samuel, Bartholomew, Erin, Hakimi, Marwa, Ye, Chun Jimmie, Bhutani, Tina, Matloubian, Mehrdad, Gensler, Lianne S., Liao, Wilson
• Format: Journal Article
• Language: English
• Published: Switzerland Frontiers Media S.A 02.03.2022
• Subjects: Arthritis, Psoriatic - diagnosis; Arthritis, Psoriatic - genetics; Biomarkers; CITE-seq; diagnostic test; Epitopes; Humans; Immunology; Machine Learning; psoriasis; Psoriasis - diagnosis; Psoriasis - genetics; psoriatic arthritis; single cell; Transcriptome; single cell; psoriatic arthritis; machine learning; diagnostic test; CITE-seq; psoriasis
• ISSN: 1664-3224; 1664-3224
• DOI: 10.3389/fimmu.2022.835760

Early diagnosis of psoriatic arthritis (PSA) is important for successful therapeutic intervention but currently remains challenging due, in part, to the scarcity of non-invasive biomarkers. In this study, we performed single cell profiling of transcriptome and cell surface protein expression to compare the peripheral blood immunocyte populations of individuals with PSA, individuals with cutaneous psoriasis (PSO) alone, and healthy individuals. We identified genes and proteins differentially expressed between PSA, PSO, and healthy subjects across 30 immune cell types and observed that some cell types, as well as specific phenotypic subsets of cells, differed in abundance between these cohorts. Cell type-specific gene and protein expression differences between PSA, PSO, and healthy groups, along with 200 previously published genetic risk factors for PSA, were further used to perform machine learning classification, with the best models achieving AUROC ≥ 0.87 when either classifying subjects among the three groups or specifically distinguishing PSA from PSO. Our findings thus expand the repertoire of gene, protein, and cellular biomarkers relevant to PSA and demonstrate the utility of machine learning-based diagnostics for this disease.